Industry 4.0 in Biomanufacturing: Predictive Real-Time Models Using Process Analytical Technology

Abstract

In biomanufacturing, process analytical technology (PAT) has become an essential tool for improving product quality, reducing costs, and increasing efficiency. In this thesis, we collect capacitance and optical density data from two in-line sensors in production bioreactors to compute real-time readings of viable cell density (VCD) and viability, two critical metrics that drive product quality and batch yield. Comparing predictions with manually collected samples, a Gaussian Process Regressor model with Matern Kernel (nu=0.5) is found to be optimal, achieving an MAPE of 7.46%, well within 10% error as defined by Amgen process development scientists. We then utilize this VCD model in conjunction with the optical density probe, which measures total cell density (TCD), in a novel way to obtain real-time measurements of viability within 5% of offline measurements conducted using a cell counter. Our results demonstrate the effectiveness of using real-time sensor data and ML models for monitoring critical quality attributes in biomanufacturing. This will enable an estimated $2M per year in savings in avoidable product losses in Amgen’s new manufacturing plant in North Carolina, approximately 50% reduction in manual sampling efforts, and offer further process improvement opportunities, particularly for advanced process control. This use case demonstrates the potential of PAT for improving biomanufacturing processes.